Target detection and parameter estimation method for sensing integrated system

Abstract

The invention relates to a target detection and parameter estimation method for an OTFS (on-the-fly-switching) integrated system, which aims at solving the problems of target detection and fractional Doppler and time delay parameter estimation of the OTFS integrated system, and comprises the following steps: step 1, a sending end generates an integrated signal, step 2, a receiver receives an echo signal and processes the echo signal, and step 3, the received echo signal is sent to a receiving end; 3, a threshold graph is generated through the integer delay and Doppler detection network based on the convolutional neural network, and target detection is carried out; and 4, fractional Doppler and delay indexes are estimated. The target detection and parameter estimation method of the OTFS communication and sensing integrated system based on deep learning has the following advantages that compared with a common CFAR detection algorithm, the detection performance of the integer time delay and Doppler detection network is better. In addition, compared with a difference algorithm which only considers integer Doppler and time delay and considers fractional Doppler and time delay, the relative distance and speed estimated by the fractional estimation network are more accurate.

Description

Technical Field

[0001] The present invention belongs to the technical field of communication perception integration, and specifically relates to a target detection and parameter estimation method of an OTFS (Orthogonal Time Frequency Space) synaesthesia integration system based on deep learning. Background Art

[0002] With the development of information technology, technical visions such as the Internet of Vehicles and the Internet of UAVs are gradually becoming a reality. These emerging application scenarios require both highly reliable communication capabilities and high-precision perception capabilities. In addition, the rapid increase in the number of communication devices has led to a growing shortage of spectrum resources. In order to solve these problems, the concept of Integrated Sensing and Communication (ISAC) was proposed. ISAC integrates the two functions of communication and perception. It can perceive the physical characteristics of the surrounding environment by actively recognizing and analyzing the characteristics of the channel during the transmission of information. Compared with systems where communication and perception are separated, it has many advantages, such as saving spectrum resources, reducing hardware costs and power consumption, and reducing device size.

[0003] In order to realize ISAC, waveform design and signal processing are very important issues. Many studies are based on orthogonal frequency division multiplexing (OFDM) to realize ISAC. However, the high Doppler effect in high-speed mobile scenarios will greatly increase the inter-carrier interference of OFDM systems, resulting in performance degradation. In this context, OTFS has attracted widespread attention because of its significant advantages in resisting time-frequency dual-domain selective fading and maintaining good communication performance in high-speed mobile scenarios. Moreover, OTFS works in the delay-Doppler domain, and the parameters of its pulse channel response are delay and Doppler. The distance and speed parameters of interest in radar perception can be derived from delay and Doppler. One of the main challenges faced by the OTFS-ISAC system is that in real scenarios, due to the limitation of delay and Doppler resolution, the actual Doppler and delay of the target are often not integer multiples of the resolution, resulting in fractional Doppler and delay errors. In addition, the number of potential targets in the perception scene is usually unknown. Therefore, how to effectively detect targets and reduce the estimation error in the presence of fractional Doppler and delay is a technical problem that needs to be solved urgently. Summary of the invention

[0004] Technical problem: In view of the shortcomings of the existing technology, the present invention provides a target detection and parameter estimation method for an OTFS synaesthesia system based on deep learning, aiming to solve the problem of target detection and fractional Doppler and delay parameter estimation of the OTFS synaesthesia system.

[0005] Technical solution: A method for target detection and parameter estimation of a joint communication and sensing integrated system according to the present invention. Considering a single-input single-output (SISO) OTFS-ISAC system of a single-base radar, the method includes the following steps:

[0006] Step 1: The transmitter generates a joint communication and sensing integrated signal. For the scenario of the base station transmitting OTFS signals, a transmission signal model is constructed.

[0007] Step 2: The receiver receives the echo signal and processes it. For the scenario of the base station receiving the echo signal, an echo signal model is constructed.

[0008] Step 3: Generate a threshold map through an integer delay and Doppler detection network based on a convolutional neural network for target detection, and solve the problems of sensing target detection and integer Doppler and delay estimation.

[0009] Step 4: Estimate the fractional Doppler and delay indices, solve the problems of fractional Doppler and delay estimation, and achieve a more refined parameter estimation effect.

[0010] Among them,

[0011] Specifically, step 1 is as follows:

[0012] 1.5. The data source generates data bits X DD ∈C NM , where C represents a complex matrix, N and M respectively represent the number of time slots and subcarriers of each OTFS frame, and their values are determined according to specific system design; the data bits are placed on the delay-Doppler plane, and the delay-Doppler plane is expressed as:

[0013]

[0014] In the formula, 1 / NT and 1 / MΔf respectively represent the sampling intervals of the Doppler axis and the delay axis, or can be regarded as the resolutions of the Doppler frequency shift and the delay; MΔf represents the bandwidth occupied by each frame, NT represents the duration of each frame; Δf is the subcarrier spacing, T is the symbol duration, and TΔf = 1 is satisfied; k and l respectively represent the indices in the Doppler domain and the delay; the values of N, M, Δf, and T are determined according to specific system design;

[0015] 1.6. Through the inverse symplectic finite Fourier transform ISFFT, X DD is transformed from the delay-Doppler domain to the time-frequency domain to obtain the time-frequency domain transmission symbol matrix X TF ∈C N×M ; X TF [n,m] ∈ X TF represents the transmission symbol at the nth time index and the mth frequency index, as shown in the following formula

[0016]

[0017] where n = 0, …, N-1, m = 0, …, M-1;

[0018] 1.7 Applying the Heisenberg transform to the transmitted symbols in the time-frequency domain can convert them into a continuous-time baseband signal s(t) as follows

[0019]

[0020] where g tx is a pulse shaping filter with a duration of T, i.e., g tx (t) = 0 for t < 0 and t > T;

[0021] 1.8 Up-converting the continuous-time baseband signal s(t) gives the time-domain transmitted signal s T (t),

[0022]

[0023] where f c is the carrier frequency.

[0024] The specific steps of step two are as follows:

[0025] 2.1 Down-converting the echo signal gives the time-domain baseband received signal r(t) as follows

[0026]

[0027] where z(t) is additive white Gaussian noise with a one-sided power spectral density PSD of N 0 , h(τ, ν) is the complex channel impulse response in the delay-Doppler domain, and τ, ν are the round-trip delay and Doppler shift, and their formula is

[0028]

[0029] where P is the number of sensed targets, and h i is the complex channel gain of the path represented by the i-th target; δ(·) represents the Dirac function; τ i and ν i represent the round-trip delay and Doppler shift of the i-th target respectively, and their formula is

[0030]

[0031] where R i is the relative distance between the i-th target and the transmitter, and V iis the relative radial velocity between the i-th target and the transmitter, and c is the speed of light; and are the delay index and Doppler index of the i-th target respectively, where l i ∈ Z and k i ∈ Z are the integer parts of the indices, and Z represents the set of integers; ι i ∈ [-0.5, 0.5] and κ i ∈ [-0.5, 0.5] are the fractional parts of the indices; when ι i = 0 and κ i = 0, that is, when both the delay and Doppler shift of the target are integer multiples of the resolution, the target is located on the lattice points of the delay-Doppler plane; when ι i ≠ 0 or κ i ≠ 0, that is, when both the delay and Doppler shift of the target are fractional multiples of the resolution, the target is located between the lattice points of the delay-Doppler plane;

[0032] 2.2. Apply the Wigner transform to the received signal to obtain the expression Y TF (t, f) in the time-frequency domain,

[0033]

[0034] where g rx (t) is the received pulse shaping filter, and A grx,r (t, f) is the cross ambiguity function between g rx (t) and the received signal, is the definition symbol, indicating that the previous formula is defined by the following formula;

[0035] 2.3. Sample Y TF (t, f) to obtain the discrete expression Y TF [n, m],

[0036]

[0037] 2.4. Use the Symplectic Finite Fourier Transform SFFT to convert Y TF [n, m] to the received symbol Y DD [k, l] in the delay-Doppler domain, where Y DD ∈ Y DD ∈ C N×M is the received symbol matrix in the delay-Doppler domain, and Y DD [k, l] is as follows

[0038]

[0039] Under the condition of an ideal pulse shaping filter, the input-output relationship between the baseband transmitted modulation symbol and the received modulation symbol is as follows

[0040] where Z DD [k, l] is additive complex Gaussian white noise in the delay-Doppler domain, with a mean of 0 and a variance of σ 2 ; h ω [k, l] is the effective channel response in the delay-Doppler domain,

[0041]

[0042]

[0043] where and are the delay index and Doppler index of the i-th target respectively, and l' and k' are variables required for calculation, representing the indices in the delay and Doppler domains;

[0044] 2.5. Perform two-dimensional pulse compression on the received signal, which essentially uses the known transmitted symbols to perform correlation operations on Y DD along the delay axis and Doppler axis respectively to obtain the matrix U; U[k, l] ∈ U is the element in the k-th row and l-th column of the matrix U, and its expression is

[0045]

[0046] where () * represents the conjugate operation, [] N and [] M represent modulo-N and modulo-M operations respectively;

[0047] 2.6. Perform a cyclic shift operation on U along the Doppler axis to make the zero axis of the Doppler index located at the center of the delay-Doppler plane;

[0048] The specific steps of step three are as follows:

[0049] 3.1. Split U into two real matrices according to the real and imaginary parts as the two inputs of the integer delay and Doppler detection network, generating the input tensor I = [U R , U I ; where U R and U I are the matrices composed of the real parts of each element in U and the matrices composed of the imaginary parts of each element respectively;

[0050] 3.2. The input first passes through two residual blocks. The residual blocks introduce residual connections, allowing the input to skip some residual blocks and enter the next part of the network. Convolutional neural networks are used in the residual blocks, and there are 5 convolutional neural network CNN layers in each residual block. Each CNN layer uses a 3*3 convolutional kernel, and the output feature maps have lengths and widths of C 1 and C 2, C 1 and C 2 is numerically equal to the number of time slots N and the number of subcarriers M. Except for the last CNN layer, the feature maps output by the remaining CNN layers need to go through group normalization GN; GN divides the channels into different groups in the channel dimension, and then calculates the mean and variance within each group to achieve normalization, thereby reducing the computational complexity of the model and accelerating convergence; the parametric rectified linear unit PReLU,

[0051]

[0052] is used for the output after normalization of the first four CNN layers to introduce non-linearity; the fifth CNN layer reduces the number of channels of the feature map to two channels, which is consistent with the input of the residual block. The output of the residual block is expressed as:

[0053] Q(x) = P(x) + x (38)

[0054] where Q(x) is the output of each residual block, including the original input x of the residual block and the output P(x) after passing through a series of layers; the ReLU function is used as the activation function at the end of each residual block, as follows

[0055] ReLU(x) = Max(0, x) (39)

[0056] In the formula, Max() represents taking the maximum value;

[0057] At the end of the network is a single convolutional layer with a kernel size of 3*3, an input channel of 2, and an output channel of 1, which reduces the dimension of the output of the residual block to form the final network output, that is, the threshold map;

[0058] 3.3. In the offline training stage, the neural network is trained through an end-to-end supervised learning method. The modulus of U is regarded as a low-resolution image composed of the target map and the threshold map, that is, |U| = T P +T N , where T P is the target map. In the target map, the amplitudes of all grid points except the grid points where the integer Doppler and delay of the target are located are 0, and T N is the threshold map. In the threshold map, only the noise amplitude is retained at the grid points where the integer Doppler and delay of the target are located, and the amplitudes of the remaining grid points are the same as |U|. T N is also the label for network training. The mean square error is used as the loss function:

[0059]

[0060] In the formula, D is the number of samples in each batch, It is the threshold map output by the network. The optimizer used during the training process is Adam, which is a gradient descent optimizer. It aims to adjust the learning rate of each parameter by calculating the first-order moment estimate and second-order moment estimate of the gradient, thereby achieving more efficient network training;

[0061] 3.4. In the online estimation stage, input I into the trained network to obtain the target detection threshold map

[0062] 3.5. Perform target detection based on the threshold map to obtain the integer Doppler and delay indices of the target;

[0063] The specific content of step four is as follows:

[0064] 4.1. Use the five data points centered on the integer target obtained in the previous step as the input data of the fractional delay and Doppler estimation network, and input the real part and the imaginary part separately;

[0065] 4.2. The fractional estimation network has a total of four layers. The first layer is the input layer with 10 neurons, and the fourth layer output layer has 1 neuron. Its output is activated by the 0.5*tanh function. The intermediate hidden layers have Q 1 and Q 2 neurons respectively, and are all activated by the tanh function. The expression of the tanh function is:

[0066]

[0067] e x is the exponential function with the real number e as the base, e -x is the negative exponential function of the real number, e≈2.71828,

[0068] 4.3. In the offline training stage, train the neural network through supervised learning. Use the Adam optimizer to adjust the network weights and biases. The mean square error is used as the loss function, and its expression is

[0069]

[0070] In the formula, T F is the true value of the fractional index, is the network estimate of the fractional index, and E is the number of samples in each batch;

[0071] 4.4. In the online estimation stage, input the real part and the imaginary part of the five data points centered on the integer target obtained in the previous step into the trained network to obtain the estimated value of the fractional index

[0072] Step five: Calculate the relative distance and relative speed. This step includes the following processes:

[0073] 5.1. Relative distance R i Calculation

[0074]

[0075] 5.2. Relative speed V i Calculation

[0076]

[0077] Beneficial effects: The target detection and parameter estimation method of the OTFS integrated communication and sensing system based on deep learning proposed in the present invention has the following advantages: Compared with the commonly used CFAR (Constant False Alarm Rate) detection algorithm, the detection performance of the integer delay and Doppler detection network is better. In addition, compared with the differential algorithms that only consider integer Doppler and delay and those that consider fractional Doppler and delay, the relative distance and speed estimated by the fractional estimation network are more accurate. Description of the drawings

[0078] Figure 1 It is the flowchart of target detection and parameter estimation of the OTFS-ISAC system based on deep learning provided by the present invention;

[0079] Figure 2 It is the integer delay and Doppler detection network structure provided by the present invention;

[0080] Figure 3 It is the fractional delay and Doppler estimation network structure provided by the present invention;

[0081] Figure 4 It is the simulation comparison diagram of detection results under different SNRs provided by the present invention; Figure 4 In (a) is the simulation comparison diagram of detection results with SNR = 5dB, Figure 4 In (b) is the simulation comparison diagram of detection results with SNR = 0dB, Figure 4 In (c) is the simulation comparison diagram of detection results with SNR = -5dB, Figure 4 In (d) is the simulation comparison diagram of detection results with SNR = -10dB.

[0082] Figure 5 It is the simulation comparison diagram of the ratio of error to resolution at different SNRs provided by the present invention. Specific implementation manners

[0083] Taking M = 32, N = 32, 4QAM modulation, carrier frequency of 24 GHz, and subcarrier spacing Δf of 39 kHz as an example, and in combination with the accompanying drawings of the specification, the specific implementation manners of the present invention will be further described in detail, such as Figure 1 、 Figure 2 、 Figure 3as shown

[0084] Step 1: The transmitting end generates a communication-sensing integrated signal, and this step includes the following processes:

[0085] (1) The data source generates data bits \(X\) DD \(\in C\) N×M , where \(C\) represents a complex matrix, \(N\) and \(M\) respectively represent the number of time slots and subcarriers of each OTFS frame, \(M = 32\), \(N = 32\). The data bits are placed on the delay-Doppler plane, and \(X\) DD [k, l]\(\in X\) DD represents the transmitted symbol at the \(k\)-th Doppler index and the \(l\)-th delay index. The delay-Doppler plane is expressed as

[0086]

[0087] In the formula, \(1 / NT\) and \(1 / M\Delta f\) respectively represent the sampling intervals of the Doppler axis and the delay axis, and can also be regarded as the resolutions of the Doppler frequency shift and the delay. \(M\Delta f\) represents the bandwidth occupied by each frame, and \(NT\) represents the duration of each frame. \(\Delta f\) is the subcarrier interval, \(T\) is the symbol duration, and \(T\Delta f = 1\) is satisfied.

[0088] (2) \(X\) is transformed from the delay-Doppler domain to the time-frequency domain through the Inverse Symplectic Finite Fourier Transform (ISFFT) to obtain the time-frequency domain transmitted symbol matrix \(X\) DD \(\in C\) TF N ×M TF , \(X\) TF [n, m]\(\in X\) tx represents the transmitted symbol at the \(n\)-th time index and the \(m\)-th frequency index, as shown in the following formula

[0089]

[0090] In the formula, \(n = 0,\cdots,N - 1\), \(m = 0,\cdots,M - 1\).

[0091] (3) Applying the Heisenberg transform to the transmitted symbol in the time-frequency domain can transform it into a continuous-time baseband signal \(s(t)\), as shown in the following formula

[0092]

[0093] In the formula, \(g\) tx is a pulse shaping filter.

[0094] (4) The continuous-time baseband signal \(s(t)\) is up-converted to obtain the time-domain transmitted signal \(s\) T (t),

[0095]

[0096] In the formula, f c is the carrier frequency.

[0097] Step 2: The receiver receives the echo signal and processes it. This step includes the following processes:

[0098] (1) Down-convert the echo signal to obtain the time-domain baseband received signal r(t), as shown in the following formula:

[0099] r(t) = ∫∫h(τ, ν)e j2πν(t-τ) s(t - τ)dτdν + z(t) (49)

[0100] In the formula, z(t) is the additive white Gaussian noise, and its one-sided power spectral density PSD is N 0 , and h(τ, ν) ∈ C is the complex channel impulse response in the delay-Doppler domain, and its formula is

[0101]

[0102] In the formula, P is the number of sensed targets, and h i is the complex channel gain of the path represented by the i-th target. δ(·) represents the Dirac function. τ i and ν i represent the round-trip delay and Doppler shift of the i-th target respectively, and their formulas are

[0103]

[0104] In the formula, R i is the relative distance between the i-th target and the transmitter, V i is the relative radial velocity between the i-th target and the transmitter, and c is the speed of light. and are the delay index and Doppler index of the i-th target respectively, where l i ∈ Z and k i ∈ Z are the integer parts of the indices, and Z represents the set of integers.

[0105] ι i ∈ [-0.5, 0.5] and κ i ∈ [-0.5, 0.5] are the fractional parts of the indices.

[0106] (2) Apply the Wigner transform to the received signal to obtain the expression Y TF (t, f),

[0107]

[0108] In the formula, g rx(t) is the received pulse shaping filter, A grx,r (t, f) is g rx The cross ambiguity function between (t) and the received signal is a defined symbol indicating that the previous formula is defined by the following formula.

[0109] (3) Sampling Y TF (t, f) can obtain the discrete expression

[0110] Y TF [n, m] = Y TF (t, f)| t＝nT,f＝mΔf (53)

[0111] (4) Using the Symplectic Finite Fourier Transform (SFFT) to convert Y TF [n, m] to obtain the received symbol Y in the delay-Doppler domain DD [k, l] ∈ Y DD , Y DD ∈ C N×M is the received symbol matrix in the delay-Doppler domain, Y DD [k, l] is as follows

[0112]

[0113] Under the condition of the ideal pulse shaping filter, the input-output relationship between the baseband transmitted modulation symbol and the received modulation symbol is as follows

[0114]

[0115] In the formula, Z DD [k, l] is the additive complex Gaussian white noise in the delay-Doppler domain, with a mean of 0 and a variance of σ 2 . h ω [k, l] is the effective channel response in the delay-Doppler domain

[0116]

[0117] In the formula is the sampling function when the same rectangular window is used at the transmitter and the receiver.

[0118]

[0119] (5) Perform two-dimensional pulse compression on the received signal, which essentially uses the known transmitted symbol to perform correlation operations on Y DD along the delay axis and the Doppler axis respectively to obtain the matrix U. U[k, l] ∈ U is the element in the k-th row and the l-th column of the matrix U, and its expression is

[0120]

[0121] Wherein, (·) * represents the conjugate operation, [·] N and [·] M represent modulo-N and modulo-M operations respectively.

[0122] (6) Perform a cyclic translation operation on U along the Doppler axis, so that the zero axis of the Doppler index is located at the center of the time delay-Doppler plane.

[0123] Step 3: Generate a threshold map through an integer time delay and Doppler target detection network based on a convolutional neural network for target detection. This step includes the following processes:

[0124] (1) Split U into two real-valued matrices according to the real part and the imaginary part as the two inputs of the integer time delay and Doppler detection network, generating an input tensor I = [U R , U I . Where U R and U I are the matrices composed of the real parts of each element in U and the matrices composed of the imaginary parts of each element respectively.

[0125] (2) Figure 2 shows the structure of the integer time delay and Doppler detection network. The input first passes through two residual blocks. The residual blocks introduce residual connections, allowing the input to skip some residual blocks and enter the next part of the network. Convolutional neural networks are used in the residual blocks, and there are 5 convolutional neural network (CNN) layers in each residual block. Each CNN layer uses a convolutional kernel of size 3*3, and the output feature maps have lengths and widths of C 1 and C 2 , C 1 and C 2 are numerically equal to the number of time slots N and the number of subcarriers M in the previous text. Except for the last CNN layer, the output feature maps of the remaining CNN layers have to go through group normalization (GN). GN divides the channels into different groups in the channel dimension, and then calculates the mean and variance within each group to achieve normalization, thereby reducing the computational amount of the model and accelerating convergence. Parametric rectified linear unit (PReLu)

[0126]

[0127] It is used for the output after normalization of the first four CNN layers to introduce non-linearity, where a = 0.25. The fifth CNN layer reduces the number of channels of the feature map to two channels, which is consistent with the input of the residual block. The output of the residual block can be expressed as

[0128] Q(x) = P(x) + x (60)

[0129] where Q(x) is the output of each residual block, including the original input x of the residual block and the output P(x) after passing through a series of layers. The ReLU function is used as the activation function at the end of each residual block, as shown in the following equation.

[0130] ReLU(x) = Max(0, x) (61)

[0131] In the equation, Max(·) represents taking the maximum value.

[0132] At the end of the network is a single convolutional layer with a kernel size of 3*3. Its input channels are 2 and the output channels are 1, reducing the dimension of the output of the residual block to form the final network output, that is, the threshold map.

[0133] (3) In the offline training stage, the neural network is trained through an end-to-end supervised learning method. Regarding the modulus of U as a low-resolution image composed of the target map and the threshold map, that is, |U| = T P +T N , where T P is the target map. In the target map, the amplitudes of all grid points except the grid points where the integer Doppler and delay of the target are located are 0. T N is the threshold map. In the threshold map, only the noise amplitude is retained at the grid points where the integer Doppler and delay of the target are located, and the amplitudes of the remaining grid points are the same as |U|. T N is also the label for network training. The mean square error is used as the loss function:[[]]

[0134]

[0135] In the equation, D is the number of samples in each batch, is the threshold map output by the network. The optimizer used during training is Adam (Adaptive Moment Estimation), which is a gradient descent optimizer designed to adjust the learning rate of each parameter by calculating the first-order moment estimate and second-order moment estimate of the gradient, thereby achieving more efficient network training.

[0136] (4) In the online estimation stage, input I into the trained network to obtain the target detection threshold map T N .

[0137] (5) Perform target detection based on the threshold map to obtain the integer Doppler and delay indices of the target.

[0138] The above steps are elaborated in detail as follows:

[0139] In the example, M = 32 and N = 32, so the size of each layer of the integer time delay and Doppler target detection network is 32 * 32. In the offline training stage, the batch size is 512, and the learning rate of the Adam optimizer is 0.001. The training set consists of 30,000 OTFS frames, where the signal-to-noise ratios SNR (Signal-to-noise Ratio) are -10 dB, -5 dB, and 0 dB respectively, with 10,000 data for each SNR. The training set focuses on the multi-target situation, and the number of targets in each OTFS frame is randomly generated from 1 to 9. The maximum target distance is 3785 meters, and the absolute value of the maximum target speed is 118 m / s. In the online estimation stage, the test set covers the situation where the SNR ranges from -10 dB to 5 dB, with an interval of 5 dB. There are 1000 test data for each SNR, and the number of targets in each OTFS frame is randomly generated from 1 to 9.

[0140] The comparison algorithm selects the commonly used CFAR detection algorithms, including CA-CFAR (Cell Averaging-CFAR), GO-CFAR (Greatest of CFAR), and SO-CFAR (Smallest of CFAR). The training interval of the three CFAR algorithms is 2 * 2, and the guard interval is 4 * 4. The detection performance is measured by the detection rate P d and the false alarm rate P fa :

[0141]

[0142] Figure 4 shows the detection results of the network and other algorithms at different SNRs with random target numbers. Obviously, regardless of the SNR, the network outperforms the CFAR algorithms. For example, in Figure 4 (b), when the false alarm rate is 10 -4 , the detection rate of the neural network is about 0.68, while the best-performing GO-CFAR among other algorithms is about 0.17. Generally speaking, as the SNR decreases, the results of all algorithms decrease, but the results of the neural network are always better than those of other algorithms.

[0143] Step 4: Estimate the fractional Doppler and time delay indices. This step includes the following processes:

[0144] (1) Take the five data points centered on the integer target obtained in the previous step as the input data of the fractional time delay and Doppler estimation network, and input the real part and the imaginary part separately.

[0145] (2) The structure of the fractional estimation network is asFigure 3 As shown, it includes four layers. The first layer is the input layer with 10 neurons. The fourth layer, the output layer, has 1 neuron, and its output is activated by the 0.5*tanh function. The middle hidden layers each have Q 1 and Q 2 neurons, and both are activated by the tanh function. The expression of the tanh function is

[0146]

[0147] (3) In the offline training stage, the neural network is trained through supervised learning, and the Adam optimizer is used to adjust the network weights and biases. The mean squared error is used as the loss function, and its expression is

[0148]

[0149] where T F is the true value of the score index, is the network estimated value of the score index, and E is the number of samples in each batch.

[0150] (4) In the online estimation stage, the real and imaginary parts of the five data points centered on the integer target obtained in the previous step are input into the trained network to obtain the estimated value of the score index

[0151] The above steps are elaborated in detail below:

[0152] Both hidden layers of the network have Q 1 = Q 2 = 32 neurons. In the offline training stage, the size of each batch is 128, and the learning rate of the Adam optimizer is 0.0001. The training set consists of 40000 data points, with 10000 data points for each of the cases where SNR = -10dB, -5dB, 0dB, and 5dB. When generating, the number of targets for each OTFS frame is also randomly generated from 1 to 9.

[0153] In the online evaluation stage, the test set covers the cases where SNR ranges from -10dB to 15dB, with an interval of 5dB. To compare the influence of different numbers of targets on the results, for each SNR, the number of targets for its OTFS frames is 4, 6, and randomly generated from 1 to 9. The maximum target distance is 3785 meters, and the absolute value of the maximum target speed is 118 m / s. There are 1000 OTFS frames for each SNR. Note that the time delay and Doppler are not distinguished here because the energy leakage patterns they cause are similar.

[0154] The comparison results in this section include the results considering only integers (assuming the integer estimation is correct, truncating and rounding) and the method based on differences. Figure 5The ratio of the root mean square error (RMSE) to the resolution at different SNRs is shown, without distinguishing between time delay and Doppler. In Figure 5 It can be clearly observed that, under different SNR conditions, the errors estimated by the fractional estimation network are significantly better than those of the method that only considers integers and the differential-based method, indicating that the network in this paper is effective for fractional estimation.

[0155] Step Five: Calculate the relative distance and relative velocity. This step includes the following processes:

[0156] (1) Relative distance R i Calculation.

[0157]

[0158] (2) Relative velocity V i Calculation.

[0159]

[0160] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Those skilled in the art to which the present invention pertains can make various modifications or supplements to the described specific embodiments or use similar ways to substitute, but will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for target detection and parameter estimation in a synaesthesia system, characterized in that: The method comprises the following steps: Step 1: The transmitter generates a synaesthesia integrated signal and builds a signal transmission model for the scenario where the base station sends an OTFS signal. Step 2: The receiver receives the echo signal and processes it. According to the scenario where the base station receives the echo signal, an echo signal model is constructed. Step 3: Target detection is performed by generating a threshold map based on an integer delay and Doppler detection network based on a convolutional neural network to solve the problems of perceived target detection and integer Doppler and delay estimation. Step 4: Estimate the fractional Doppler and delay indexes to solve the fractional Doppler and delay estimation problems and achieve a more refined parameter estimation effect.

2. The method for target detection and parameter estimation of a synaesthesia system according to claim 1, characterized in that: The step 1 is specifically as follows: 1.

1. The source generates data bits X DD ∈C NM , C represents a complex matrix, N and M represent the number of time slots and subcarriers in each OTFS frame, respectively, and their values ​​are determined according to the specific system design; the data bits are placed on the delay-Doppler plane, and the delay-Doppler plane is expressed as: Where, 1 / NT and 1 / MΔf represent the sampling intervals of the Doppler axis and the delay axis, respectively, or can be regarded as the resolution of Doppler frequency shift and delay; MΔf represents the bandwidth occupied by each frame, and NT represents the duration of each frame; Δf is the subcarrier spacing, T is the symbol duration, and satisfies TΔf=1; k and l represent the indexes of the Doppler domain and the delay respectively; the values ​​of N, M, Δf and T are determined according to the specific system design; 1.

2. Transform X by symplectic inverse Fourier transform ISFFT DD Convert from the delay-Doppler domain to the time-frequency domain to obtain the time-frequency domain transmit symbol matrix X TF ∈C N×M ; X TF [n,m]∈X TF Represents the transmitted symbol at the nth time index and the mth frequency index, as follows Where n = 0, ..., N-1, m = 0, ..., M-1; 1.

3. Applying the Heisenberg transform to the transmitted symbol in the time-frequency domain can convert it into a continuous-time baseband signal s(t), as shown below: In the formula, g tx It is a pulse shaping filter with a duration of T, i.e. g tx (t) = 0, t < 0, t > T; 1.

4. Up-convert the continuous-time baseband signal s(t) to obtain the time-domain transmitted signal s T (t), In the formula, f c is the carrier frequency.

3. The method for target detection and parameter estimation of a synaesthesia system according to claim 2, characterized in that: The step 2 is specifically as follows: 2.

1. Down-convert the echo signal to obtain the time domain baseband received signal r(t), as shown below r(t)=∫∫h(τ,ν)e j2πν(t-τ) s(t-τ)dτdν+z(t) (5) Where z(t) is additive white Gaussian noise, whose single-sided power spectral density PSD is N0, h(τ, ν) is the complex channel impulse response in the delay-Doppler domain, τ, ν are the round-trip delay and Doppler frequency shift, and the formula is: Where P is the number of perceived targets, h i is the complex channel gain of the path represented by the i-th target; δ(·) represents the Dirac function; τ i and ν i Represent the round-trip delay and Doppler frequency shift of the i-th target respectively, and the formula is: In the formula, R i is the relative distance between the i-th target and the transmitter, V i is the relative radial velocity between the ith target and the transmitter, c is the speed of light; and are the delay index and Doppler index of the i-th target, respectively, where l i ∈Z and k i ∈Z is the integer part of the index, Z represents the set of integers; ι i ∈[-0.5,0.5] and κ i ∈[-0.5,0.5] is the fractional part of the index; when ι i = 0 and κ i = 0, that is, when the target's delay and Doppler shift are both integer multiples of the resolution, the target is located on the grid point of the delay-Doppler plane; when ι i ≠0 or κ i When ≠0, that is, when the target's delay and Doppler shift are both fractional times of the resolution, the target is located between the grid points of the delay-Doppler plane; 2.

2. Applying Wigner transform to the received signal yields the expression Y in the time-frequency domain: TF (t,f), In the formula, g rx (t) is the receive pulse shaping filter, A grx,r (t,f) is g rx (t) and the cross ambiguity function between the received signal, is a definition symbol, indicating that the preceding formula is defined by the following formula; 2.

3. For Y TF (t,f) is sampled to obtain the discrete expression Y TF [n,m], Y TF [n,m]=Y TF (t,f)| t＝nT,f＝mΔf (9) 2.

4. Use the symplectic finite Fourier transform SFFT to transform Y TF [n,m] is converted into the received symbol Y in the delay-Doppler domain DD [k,l]∈Y DD , Y DD ∈C N×M is the delay-Doppler domain received symbol matrix, Y DD [k,l] is as follows Under ideal pulse shaping filter conditions, the input and output relationship between the baseband transmission modulation symbol and the received modulation symbol is as follows In the formula, Z DD [k,l] is an additive complex Gaussian white noise in the delay-Doppler domain with a mean of 0 and a variance of σ 2 ;h ω [k,l] is the effective channel response in the delay-Doppler domain, In the formula, The sampling function when the transmitting end and the receiving end use the same rectangular window; In the formula, and are the delay index and Doppler index of the ith target, respectively. l′ and k′ are variables required for calculation, representing the index in the delay and Doppler domains. 2.

5. Perform two-dimensional pulse compression on the received signal. In essence, it uses the known transmitted symbols to compress the Y axis along the delay axis and the Doppler axis. DD Do the relevant operations to get the matrix U; U[k,l]∈U is the element in the kth row and lth column of the matrix U, and its expression is In the formula, () * represents the conjugate operation, [] N and[] M Respectively represent modulo N and modulo M operations; 2.

6. Perform a circular translation operation on U along the Doppler axis so that the Doppler index zero axis is located in the center of the delay-Doppler plane.

4. The method for target detection and parameter estimation of a synaesthesia system according to claim 3, characterized in that: The step three is specifically as follows: 3.

1. Split U into two real matrices according to the real part and the imaginary part as two inputs of the integer delay and Doppler detection network, generating the input tensor I = [U R ,U I ]; where U R and U I are the matrices consisting of the real part of each element in U and the matrices consisting of the imaginary part of each element; 3.

2. The input first passes through two residual blocks. The residual blocks introduce residual connections, allowing the input to skip certain residual blocks and enter the next part of the network. Convolutional neural networks are used in the residual blocks, and each residual block has 5 convolutional neural network CNN layers. Each CNN layer uses a 3*3 convolution kernel, and the length and width of its output feature map are C1 and C2 respectively. C1 and C2 are numerically equal to the number of time slots N and the number of subcarriers M. Except for the last CNN layer, the feature maps output by the remaining CNN layers must undergo group normalization GN; GN divides the channels into different groups in the channel dimension, and then calculates the mean and variance within each group to achieve normalization, thereby reducing the amount of calculation of the model and accelerating convergence; parameter corrected linear unit PReLU, It is used for the normalized output of the first four CNN layers to introduce nonlinearity; the fifth CNN layer reduces the number of channels of the feature map to two channels, which is consistent with the input of the residual block. The output of the residual block is expressed as: Q(x)=P(x)+x (16) Where Q(x) is the output of each residual block, including the original input x of the residual block and the output P(x) after a series of layers; the ReLU function is used as the activation function at the end of each residual block, as shown below ReLU(x)=Max(0,x) (17) In the formula, Max() means taking the maximum value; At the end of the network is a single convolution layer with a convolution kernel size of 3*3, 2 input channels, and 1 output channel, which reduces the dimension of the output of the residual block and forms the final network output, i.e., the threshold map. 3.

3. In the offline training phase, the neural network is trained through an end-to-end supervised learning method, and the model of U is regarded as a low-resolution image composed of the target image and the threshold image, that is, |U| = T P T N , where T P is the target graph. In the target graph, except for the grid points where the integer Doppler and delay of the target are located, the amplitudes of the other grid points are all 0. N is the threshold map. In the threshold map, only the noise amplitude is retained at the grid points where the integer Doppler and delay of the target are located. The amplitudes of the remaining grid points are the same as |U|. N It is also the label for network training, using mean square error as the loss function: In the formula, D is the number of samples in each batch, is the threshold map of the network output. The optimizer used in the training process is Adam, which is a gradient descent optimizer that aims to adjust the learning rate of each parameter by calculating the first-order moment estimate and the second-order moment estimate of the gradient, thereby achieving more efficient network training. 3.

4. In the online estimation stage, I is input into the trained network to obtain the target detection threshold map 3.

5. Perform target detection based on the threshold map and obtain the integer Doppler and delay index of the target.

5. The method for target detection and parameter estimation of a synaesthesia system according to claim 4, characterized in that: The step 4 is specifically as follows: 4.

1. Use the five data points centered on the integer target obtained in the previous step as the input data of the fractional delay and Doppler estimation network, and input the real part and the imaginary part separately; 4.

2. The score estimation network has four layers. The first layer is the input layer, which has 10 neurons. The fourth layer is the output layer, which has 1 neuron. Its output is activated by 0.5*tanh function. The middle hidden layer has Q1 and Q2 neurons, which are activated by tanh function. The expression of tanh function is: e x is an exponential function with the real number e as its base, e -x Taking the negative exponential function of real numbers, e≈2.71828, 4.

3. During the offline training phase, the neural network is trained by supervised learning. The Adam optimizer is used to adjust the network weights and biases. The mean square error is used as the loss function, which is expressed as Where, T F is the true value of the fractional index, is the network estimate of the score index, and E is the number of samples in each batch; 4.

4. In the online estimation stage, the real and imaginary parts of the five data points centered on the integer target obtained in the previous step are input into the trained network to obtain the estimated value of the fractional index Step 5: Calculate relative distance and relative speed. This step includes the following process: 5.

1. Relative distance R i calculate 5.

2. Relative velocity V i calculate

Images